Introducing Clairic, a new System One class of extraction models
blog.cloudflare.com
October 19, 2026
21 min read
45/100
Summary
Cloudflare has released Clairic in preview on Workers AI as @cf/cloudflare/clairic, introducing a System One extraction model designed to return typed values from unstructured text and up to four images. The company plans to release open weights later. Clairic can extract source spans, names, dates, numbers and lists; compute values such as totals or date differences; estimate bounded values with medians and 80% intervals; and forecast time-series values. It processes requests in a single forward pass without generating output tokens, returning typed values or null along with confidence scores and, for scalar fields, a probability that the value is absent. Cloudflare says Clairic is intended for tasks where generative LLM output requires parsing and validation, including invoice processing, receipt reading and support-ticket triage. In a 21-day test forecasting daily TSA checkpoint volumes, Cloudflare reported a 3.3% average error and said 17 of 21 actual counts fell within Clairic’s 80% prediction intervals. A damage-claim workflow combining Clairic, Cloudflare’s Clef decision model and GLM-5.3 took 19.3 seconds and cost $6,032 per million claims, compared with 61.7 seconds and $11,740 for an all-GLM-5.3 workflow, according to Cloudflare. The company cautions that longer arithmetic chains, unfamiliar estimates and detailed long-horizon forecasts remain limitations.
Key Takeaways
- Clairic is available in preview through Cloudflare Workers AI under the model name @cf/cloudflare/clairic, with an open-weights release planned later.
- Clairic returns typed extracted, computed, estimated and forecast values, including confidence measures, null probabilities and 80% intervals for estimates and forecasts.
- Cloudflare says Clairic uses one parallel forward pass and no output tokens, allowing multiple fields to be processed from the same input without generative text output.
- Cloudflare reported that Clairic’s 21 daily TSA-passenger forecasts had a 3.3% average error, with 17 of 21 observed values inside its 80% intervals.
- Cloudflare cautions that Clairic is not a replacement for dedicated time-series systems and currently struggles with long multi-step arithmetic and estimates lacking supporting evidence.
What the discussion said
The conversation focused less on the launch rhetoric and more on what a compact extraction model could unlock in practical AI systems. Several commenters welcomed the move away from one giant general-purpose model toward small specialists that can classify, extract fields, and make narrow decisions cheaply. The appeal is operational as much as technical: readers imagine these models embedded throughout software delivery workflows, where lightweight containerized inference could turn currently expensive or brittle AI checks into routine infrastructure. One participant also saw likely usefulness beyond extraction proper, especially for named-entity recognition and intent/slot parsing. Enthusiasm is tempered by a demand for evidence. The most pointed criticism is that the release offers no published benchmarks, leaving readers unable to judge whether the System One framing represents a meaningful modeling advance, a better packaging of established extraction techniques, or simply a new label. Interest in open weights also surfaced immediately, suggesting that deployability and independent evaluation will strongly shape whether the model gains traction. Overall, the thread is optimistic about task-specialized small models, but treats Clairic's claimed distinctiveness as unproven until performance data arrives.
Where opinion split
The sharpest dispute is whether System One denotes a genuinely new extraction-model approach or a rebrand of familiar compact classifiers and extractors. Optimists see focused, inexpensive models as the practical path to deploying AI everywhere; skeptics argue that without benchmarks or technical detail, there is no basis to distinguish this from prior art.
Coverage & discussions
Hacker News · blog.cloudflare.com
Community discussions
Community Sentiment
Positives
- Specialized extraction models could make AI a cheap, dependable building block for narrow workflow decisions instead of forcing every task through a large generalist model.
- Small models that fit naturally into ubiquitous containerized CI/CD environments could bring classification and extraction checks into ordinary development pipelines at negligible marginal cost.
- Readers expect the approach to extend well to named-entity recognition and intent-slot parsing, where constrained outputs reward focused models over broad chat behavior.
- Demand for open weights reflects real enthusiasm for locally deployable extraction AI that teams can inspect, adapt, and validate independently.
Concerns
- The absence of published benchmarks makes the launch impossible to evaluate: readers cannot tell whether claimed extraction gains are real, competitive, or merely asserted.
- The System One label risks obscuring familiar prior art unless the creators show a concrete methodological difference and evidence that it improves outcomes.
